Thanks for the response. I ***think*** I'm making a bit of
progress ....
On 29/07/2008, at 10:14 AM, Douglas Bates wrote:
On Sun, Jul 27, 2008 at 9:06 PM, Rolf Turner
<[EMAIL PROTECTED]> wrote:
<snip>
What I *don't* understand is the correlation structure of the
estimates
produced by lmer(), which is:
Correlation of Fixed Effects:
(Intr) tstnm2 tstnm3 tstnm4 tstnm5
tstnum2 -0.434
tstnum3 -0.434 0.500
tstnum4 -0.434 0.500 0.500
tstnum5 -0.434 0.500 0.500 0.500
tstnum6 -0.434 0.500 0.500 0.500 0.500
So apparently the way I called lmer() places substantial constraints
on the covariance structure.
That's the correlation matrix of the fixed-effects parameters. You
should have separately gotten estimates of the variance-covariance of
the random effects, which you coyly did not show us :-). Because you
are allowing only a simple, scalar random effect per student there
will be an estimate of the variance of this random effect and an
estimate of the residual variance.
Far be from me to be coy! The thing is, I'm floundering
in the dark, not really understanding what I'm doing.
Here is the bit of the output of summary() that I didn't show:
Data: schooldat
AIC BIC logLik MLdeviance REMLdeviance
4470 4507 -2228 4435 4456
Random effects:
Groups Name Variance Std.Dev.
stdnt (Intercept) 1.41915 1.19128
Residual 0.85903 0.92684
number of obs: 1440, groups: stdnt, 240
How can I (is there any way that I can)
tell lmer() to fit the most general possible covariance structure?
It sounds like you want a model formula of
lmer(y ~ tstnum + (0 + tstnum|stdnt), data=schooldat)
but that model will have 21 variance-covariance terms to estimate (22
if you count the residual variance but that one gets profiled out of
the optimization). I would not be surprised if the estimated
variance-covariance matrix for the random effects turns out to be
singular.
Tried that; got a warning message:
Warning message:
In .local(x, ..., value) : nlminb returned message false convergence (8)
I need to mess around with the output produced and see if I can fit
it into my mental structures. I think I'm converging on comprehension
of the foregoing syntax, but.
If anyone wishes to experiment with the real data set (it's a bit
too big to post here) I can make it available to them via email.
Generally I would jump at the chance but not with my "To Do" list in
its current, sadly over-committed, state.
Pity, but perfectly understandable. Any other takers? :-)
Thanks again.
cheers,
Rolf
######################################################################
Attention:\ This e-mail message is privileged and confid...{{dropped:9}}
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
